A Web Service for Pre-segmenting Very Long Transcribed Speech Recordings

A Web Service for Pre-segmenting Very Long Transcribed Speech Recordings Nina Poerner∗, Florian Schiel† ∗Center for Information and Language Processing, University of Munich, Germany †Bavarian Archive for Speech Signals, University of Munich, Germany poerner@cis.uni-muenchen.de, schiel@bas.uni-muenchen.de Abstract The run time of classical text-to-speech alignment algorithms tends to grow quadratically with the length of the input. This makes it difficult to apply them to very long speech recordings. In this paper, we describe and evaluate two algorithms that pre-segment long recordings into manageable “chunks”. The first algorithm is fast but cannot guarantee short chunks on noisy recordings or erroneous transcriptions. The second algorithm reliably delivers short chunks but is less effective in terms of run time and chunk boundary accuracy. We show that both algorithms reduce the run time of the MAUS speech segmentation system to under real-time, even on recordings that could not previously be processed. Evaluation on real-world recordings in three different languages shows that the majority of chunk boundaries obtained with the proposed methods deviate less than 100 ms from a ground truth segmentation. On a separate German studio quality recording, MAUS word segmentation accuracy was slightly improved by both algorithms. The chunking service is freely accessible via a web API in the CLARIN infrastructure, and currently supports 33 languages and dialects. Keywords: text-to-speech alignment, speech segmentation, speech processing, MAUS 1. Introduction Text-to-speech alignment plays an important role in Pho- netic research, speech technology, and in the production of video subtitles. The Bavarian Archive for Speech Signals offers a free web service, Munich AUtomatic Segmenta- tion (MAUS) (Schiel, 1999), to the scientific community. MAUS segments transcribed speech recordings into words and phones. Its main advantage is that, for many of its 33 languages and dialects, it cannot only perform forced align- ment, but also model variable pronunciation. MAUS has recently been evaluated as the best aligner, out of ten can- didates, on a recording by J.F. Kennedy (Oesch and Sidler, 2017). One major disadvantage of MAUS is its run time on long recordings. This is because MAUS performs optimal Viterbi alignment, which, while more accurate than beam search, has quadratic run time with signal duration and tran- scription length (see Figure 1). Long run times mainly af- fect users who wish to use MAUS on certain types of real- world recordings, such as public speeches, interviews or audio books, as these tend to be longer than audio material produced for scientific research. One way to combat MAUS’s run time problem is to pre-segment long recording-transcription pairs into short “chunks”, i.e., slices of matching audio and text material of a few seconds or minutes. Recently, we introduced the “chunker”, an algorithm that does this job automatically (Poerner and Schiel, 2016). The chunker has since been made available as a web service by the Bavarian Archive for Speech Signals (Kisler et al., 2017). In the following, we give an overview of the algorithm, in- cluding a number of new developments (Section 2.). Sec- tions 3. and 4. describe an evaluation on …

Content truncated to 500 words